Latest / The CEO Diary with Fexingo: Leadership Lessons, Executive Decisions, and Corner Office Stories / How Nvidia Reinvented Its Business Model Twice
Transcript
- Lucas: So today, the most valuable semiconductor company in history wasn't built on the product you think it was. Luna: You mean Nvidia wasn't always about AI data centers? Lucas: Exactly. In 2006, Nvidia was a gaming graphics company. Pretty good one, sure — but their entire revenue came from selling chips that made video games look better. Then Jensen Huang made a bet that most analysts thought was nuts. Luna: The CUDA bet. Lucas: Right. CUDA — Compute Unified Device Architecture. It was a software layer that let developers use the GPU for general-purpose computing, not just graphics. The problem? At launch, it had zero applications. Nobody needed it. And Nvidia was spending enormous engineering dollars building hardware that supported this software that nobody was using. Luna: How much were they spending? Lucas: Rough estimates put it at hundreds of millions of dollars annually in the early years. For a company whose revenue in 2006 was around two and a half billion, that was a huge bet. And for years, it didn't pay off. Wall Street hated it. Competitors like AMD and Intel focused on simpler architectures. But Huang kept going. Luna: So what changed? When did CUDA become the thing? Lucas: Two inflection points. First, in 2009, researchers at Stanford showed that Nvidia GPUs running CUDA could accelerate the processing of deep learning neural networks by 70 times compared to CPUs. Suddenly, the AI research community had a tool. Then, around 2012, Alex Krizhevsky won the ImageNet competition using two Nvidia GPUs. That was the moment deep learning went mainstream, and CUDA was the only platform that could run it. Luna: So it took six years from launch to the first real use case. Lucas: Yeah. Six years of negative ROI, internal skepticism, and external doubt. That first reinvention was not from graphics to AI — it was from selling hardware to building a platform. Because CUDA locked developers into Nvidia's ecosystem. Once a researcher wrote their code in CUDA, they couldn't easily switch to an AMD card. That's the platform moat. Luna: Right, it's not just about the chip. It's about the software that makes the chip indispensable. Lucas: Exactly. And that was the first reinvention. But there's a second one, and it's even more recent. In 2020, Nvidia announced the acquisition of Mellanox for about seven billion dollars. Mellanox made high-speed networking gear for data centers. At the time, people thought, 'why is a GPU company buying a networking company?' Luna: I remember that. The deal seemed weird. Lucas: It was weird until you understood Huang's vision. He realized that AI training was becoming a system problem, not a chip problem. You can't just throw more GPUs at a model — you need them to communicate at lightning speed. Mellanox gave Nvidia the networking piece. Then they started building complete server systems, like the DGX line, which integrated GPUs, networking, and cooling into a single box. Luna: So they went from selling components to selling entire data center solutions? Lucas: Yes. By 2024, Nvidia's data center revenue was over forty billion dollars — more than ten times their gaming revenue. They were no longer a graphics company. They were a systems company. And the interesting thing is, this second reinvention was about vertical integration, not just platform lock-in. Luna: How did they manage to execute two massive shifts without imploding? Most companies fail at one. Lucas: Huang's leadership style is a big part of it. He's known for having a flat organization — there are something like sixty direct reports to him. He calls it the 'one-team approach.' No layers of management to slow down decisions. He also pushes a culture of 'intellectual honesty' where bad news travels fast. Luna: But that flat structure sounds chaotic. How does it work? Lucas: It works because Huang is incredibly deep technically. He can evaluate the details himself, so he doesn't need layers of VPs to filter information. But it's also exhausting. A former executive described working with him as 'like drinking from a firehose of genius.' Not everyone can handle it. Luna: So the lesson for other CEOs might be: bet big on an unproven market, and be willing to wait half a decade? Lucas: That's part of it. But the deeper lesson is about business model flexibility. Nvidia didn't just launch a product. They changed what kind of company they were — twice. From component maker to platform, then from platform to full-systems integrator. Each time, they had to build capabilities that didn't exist inside the company. Luna: And they did it without destroying their existing business. The gaming segment still made money while they were investing in data center. Lucas: Right. They used the profits from gaming to fund the AI bet. That's classic 'ambidextrous organization' thinking — exploit the current business while exploring the future. But it requires discipline. Most companies either milk the cash cow until it dies, or invest so heavily in the new thing that the old thing collapses. Luna: What about the competitive response? Intel tried to build its own GPU, AMD has its own platform. None have really dented Nvidia. Lucas: That's the power of the platform moat combined with the systems moat. Even if Intel builds a comparable GPU, they would need to also build the networking, the software stack, the ecosystem. Nvidia has years of head start on all three. And Huang keeps reinvesting — in fiscal year 2026, they spent over ten billion dollars on R&D. Luna: Ten billion. That's more than Intel's entire R&D budget in some years. Lucas: It's staggering. And it's a signal that they intend to stay ahead. But there's a risk too. If the AI market shifts — say, if inference moves to specialized chips that don't need Nvidia's full stack — they could be overbuilt. But that's the gamble. Luna: Before we go further, Lucas, I want to say something. This kind of deep-dive conversation — it's part of why I love doing this show. Lucas: Yeah, me too. And if it's worth a coffee to you, there's a link. It's buy me a coffee dot com slash fexingo. Smallest possible ask. It keeps this ad-free and focused. Luna: Honestly, every single contribution adds up. And it's easy to skip if it's not your thing. Lucas: Alright, back to Nvidia. So we talked about the two reinventions. But there's a third dimension that's often overlooked: how Nvidia managed its relationship with its customers during these shifts. Luna: You mean the car companies and cloud providers that are also competitors? Lucas: Exactly. In the autonomous driving space, Nvidia supplies chips to Toyota, Mercedes, and others, while also developing its own full-stack solutions. Some car makers worry about dependency. Similarly, Amazon and Google use Nvidia chips but also build their own custom AI accelerators. Nvidia has to walk a tightrope: be indispensable but not threatening. Luna: How do they handle that? Lucas: Huang's approach is to be transparent about their roadmap and to emphasize that they are a platform company, not a competitor to their customers. They offer customization options. But it's an ongoing tension. For instance, in 2024, Microsoft and Meta were among Nvidia's biggest customers, but both are investing in custom silicon. It's a delicate dance. Luna: So the next reinvention might be how Nvidia positions itself in a world where its biggest customers become its rivals. Lucas: That's the question. And Huang is already hinting at it. He talks about Nvidia as an 'AI factory' — a platform that any company can use to produce intelligence. That's a services play, not just hardware. If they can successfully sell AI infrastructure as a service, it changes the game again. Luna: So we might be seeing a third reinvention in real time? Lucas: Maybe. And if history is any guide, it'll take five years, cost billions, and look like a terrible idea at first. Then it'll work. Luna: And that's exactly why we cover these stories.